In the architecture of modern Natural Language Processing (NLP), tasks like coreference resolution, entity linking, and event extraction are typically framed as multi-stage pipelines. First, a system must identify the relevant spans of text—the "mentions" or "triggers." Second, it must resolve or link these spans to entities, clusters, or knowledge bases.
A recurring finding across decades of research is that the first stage is not merely a precursor; it is a critical bottleneck. If a system fails to detect a mention at the outset, no amount of sophisticated reasoning in the second stage can recover it. This "mention detection bottleneck" creates a performance ceiling that constrains the entire pipeline.
The Unrecoverable Error
The intuition behind this bottleneck is simple: you cannot link what you have not found. In BESTCUT: A Graph Algorithm for Coreference Resolution, Cristina Nicolae and Gabriel Nicolae noted that errors in identifying mentions are typically unrecoverable. This dependency creates a pipeline where downstream clustering cannot compensate for upstream extraction failures.
This is particularly visible in "zero-shot" or "out-of-the-box" scenarios. Even as Large Language Models (LLMs) show immense promise, Alan Ritter and Nghia T. Le found in Are Large Language Models Robust Zero-shot Coreference Resolvers? that identifying specific spans—especially nominal phrases—remains a significant hurdle. Without accurate mention detection, the subsequent task of linking mentions into coherent clusters remains highly error-prone.
Detection as the Primary Driver of Progress
For years, the NLP community focused on refining the logic used to link mentions (e.g., better clustering algorithms or more features). However, Vincent Ng argued in Conundrums in Entity Coreference Resolution: Making Sense of the State of the Art that recent performance gains are primarily driven by improved mention detection rather than more sophisticated linking logic.
When models are provided with "gold" (perfect) mention boundaries, their accuracy increases significantly. This suggests that the ability to identify candidate spans is a more critical bottleneck than the reasoning required to link them. Altaf Rahman and Vincent Ng demonstrated in Supervised Models for Coreference Resolution that noise introduced during the automatic extraction of noun phrases creates a performance ceiling that subsequent algorithms simply cannot overcome.
A Universal Pattern Across Tasks
This bottleneck is not limited to standard coreference resolution; it appears wherever a system must first "flag" an item before processing it:
- Event Extraction: In the Overview of TAC KBP 2015 Event Nugget Track, Teruko Mitamura highlighted that the accuracy of event coreference is strictly bounded by the performance of the preceding detection stage. If a system fails to identify an initial event mention (the "trigger"), it is impossible to establish subsequent links.
- Zero Anaphora: When resolving omitted subjects in languages like Chinese or Japanese, the primary source of failure is often the initial discovery of the "gap" itself. Fang Kong and Guodong Zhou showed in A Tree Kernel-Based Unified Framework for Chinese Zero Anaphora Resolution that improving the sensitivity of gap detection is more critical for overall performance than refining the identification of the missing referents.
- Spelling Correction: Even in character-level tasks, Xiao Wei observed in PTCSpell: Pre-trained Corrector Based on Character Shape and Pinyin for Chinese Spelling Correction that the correction module is strictly bounded by the sensitivity of the initial detection stage. If the detector fails to flag an error, the corrector never receives the opportunity to fix it.
The Limits of End-to-End Modeling
To combat the rigid nature of pipelines, researchers have moved toward "end-to-end" neural models that jointly learn detection and resolution. However, the bottleneck persists.
In End-to-End Neural Event Coreference Resolution, Hongyu Lin and Le Sun found that even when modeling detection and coreference jointly, errors in identifying mentions remain a primary source of failure. Similarly, Wei Wu and colleagues noted in CorefQA: Coreference Resolution as Query-based Span Prediction that while their QA-based approach can "recover" some mentions, the system still relies on an initial proposal module to trigger the search. If this module fails to detect every mention within a specific entity cluster, the entire entity remains unresolvable.
Why Intuition Matters
The core intuition for researchers is that recall in the detection phase is paramount. While precision is important to avoid "hallucinating" entities, a drop in detection recall is a fatal blow to the rest of the system. As Ting Liu and Weipeng Chen summarized in their CoNLL 2011 Shared Task paper, the system cannot link entities that were never correctly identified as mentions in the first place. Improving the upstream identification of entity boundaries is almost always more impactful than refining the downstream linking logic.